"The world is one big data problem.”
Introducing UncertaintiesThe accuracy of LeNet-5 decreases drastically when tested on digits that are shifted and overlaid with colored backgrounds.
To tackle this, we can:
...which is implemented by:
|
|
ResultsWhen LeNet-5 is applied to digits with random colored background and shifts, it yielded a mAP of 80.18%. This value is quite low compared to the mAP of 99.14% when it was applied to the original test set. This show that the LeNet-5 is not distortion-invariant.
On the other hand, the modified CNN yielded more desirable results. This suggests that increasing the depth helps the neural net extract more salient features; while overlapped maxpooling helps the network to become more distortion-invariant. |
|